Error Estimation in Astronomy: A Guide

Error estimation in astronomy: A guide
René Andrae
Max-Planck-Institut für Astronomie, Königstuhl 17, 69117 Heidelberg, Germany
e-mail: andrae@mpia-hd.mpg.de
Estimating errors is a crucial part of any scientific analysis. Whenever a parameter is

estimated (model-based or not), an error estimate is necessary. Any parameter estimate
that is given without an error estimate is meaningless. Nevertheless, many (undergraduate
arXiv:1009.2755v3 [astro-ph.IM] 29 Oct 2010
or graduate) students have to teach such methods for error estimation to themselves when
working scientifically for the first time. This manuscript presents an easy-to-understand
overview of different methods for error estimation that are applicable to both model-based
and model-independent parameter estimates. These methods are not discussed in detail,
but their basics are briefly outlined and their assumptions carefully noted. In particular,
the methods for error estimation discussed are grid search, varying χ2 , the Fisher matrix,
Monte-Carlo methods, error propagation, data resampling, and bootstrapping. Finally, a
method is outlined how to propagate measurement errors through complex data-reduction
pipelines.
1 Introduction
This manuscript is intended as a guide to error estimation for parameter estimates in astron-
omy. I try to explain several different approaches to this problem, where the emphasis is on
highlighting the diversity of approaches and their individual assumptions. Making those as-
sumptions explicitly clear is one of the major objectives, because using a certain method in a
situation where its assumptions are not satisfied will result in incorrect error estimates. As this
manuscript is just an overview, the list of methods presented is by no means complete.
A typical task in scientific research is to make measurements of certain data and then to
draw inferences from them. Usually, the inference is not drawn directly from the data but
rather from one or more parameters that are estimated from the data. Here are two examples:
• Apparent magnitude of stars or galaxies. Based on a photometric image, we need to
estimate the parameter “flux” of the desired object, before we can infer its apparent
magnitude.
• Radial velocity of stars. First, we need to take a spectrum of the star and identify appro-
priate emission/absorption lines. We can then estimate the parameter “radial velocity”
from fitting these spectral lines.
Whenever such parameter estimates are involved, it is also crucial to estimate the error of the
resulting parameter.
What does a parameter estimate and its error actually signify? More simply, what is the
meaning of an expression such as 4.3 ± 0.7? This question will be answered in detail in Sects.
2.3 and 2.5, but we want to give a preliminary answer here for the sake of motivation. The
crucial point is that the true result of some parameter estimate is not something like 4.3 ± 0.7,
but rather a probability distribution for all possible values of this parameter. An expression
like 4.3 ± 0.7 is nothing more than an attempt to encode the information contained in this
1
René Andrae (2010) – Error estimation in astronomy: A guide
probability distribution in a more simple way, where the details of this “encoding” are given
by some general standards (cf. Sect. 2.5). Put simply, the value 4.3 signifies the maximum of
the probability distribution (the most likely value), whereas the “error” 0.7 signifies the width
of the distribution. Hence, the value 4.3 alone contains insufficient information, since it does
not enable us to reconstruct the probability distribution (the true result). More drastically:
A parameter value without a corresponding error estimate is meaningless. Therefore, error
estimation is equally as important an ingredient in scientific work as parameter estimation itself.
Unfortunately, a profound and compulsory statistical eduction is missing in many university
curricula. Consequently, when (undergraduate or graduate) students are faced with these
problems for the first time during their research, they need to teach it to themselves. This is
often not very efficient and usually the student focuses on a certain method but does not gain
a broader overview. The author’s motivation was to support this process by providing such an
overview.
Where do uncertainties stem from? Of course, an important source of uncertainties is the
measured dataset itself, but models can also give rise to uncertainties. Some general origins of
uncertainties are:
• Random errors during the measurement process.
• Systematic errors during the measurement process.
• Systematic errors introduced by a model.
We obviously have to differentiate between random and systematic errors, i.e., between vari-
ance/scatter and bias/offset. Systematic errors are usually very hard to identify and to correct
for. However, this is not part of this manuscript, since individual solutions strongly depend on
the specific problem. Here, different methods of estimating random errors (variance/scatter)
are considered, i.e., those quantities that determine the size of error bars or, more generally,
error contours. Error estimation for parameter estimation only is described, whereas error
estimation for classification problems is not discussed.
This manuscript will not be submitted to any journal for two reasons: First, its content
is not genuinely new but a compilation of existing methods. Second, its subject is statistical
methodology rather than astronomy. Any comments that may improve this manuscript are
explicitly welcome.
2 Preliminaries
Before diving into the different methods for error estimation, some preliminaries should be
briefly discussed, firstly, the terminology used throughout this manuscript and, secondly, errors
of data. Thirdly, the basics of parameter estimation are briefly explained, including the intro-
duction of the concept of a likelihood function. Fourthly, the central-limit theorem is discussed.
Finally, the concept of confidence intervals, which are the desired error estimates, is introduced.
2.1 Terminology
This manuscript is about “error estimation for parameter estimates”. The first step is usually
to measure some data and also to measure its error or uncertainty (Sect. 2.2). Given this
measurement, the task is then to estimate some parameter (Sect. 2.3). The estimate of the
parameter θ is denoted by a hat, θ̂, which is common practice in statistics. Here, I want
to introduce the concept of a qualitative difference between “measuring” and “estimating”:
Measurements are outcomes of a real experiment. Conversely, estimates are inferences from
2
Figure 1: Examples of Poisson distribu- Figure 2: Poisson distribution with µ = 10

tions with µ = 1, µ = 2, µ = 4, and µ = 8. and Gaussian distribution with σ 2 = µ =
10. The Gaussian is a very good approxi-
mation to the Poisson distribution.
measurements, i.e., they are not directly related to experiments. Although this difference is
not of vital importance, both terms are rigorously differentiated throughout this manuscript in
order to make clear what is being referred to.
Another issue of terminology concerns the words “error” and “uncertainty”. As mentioned
in the introduction, systematic errors are not considered here, and hence both terms may
be used more or less synonymously.1 There is also a third word of the same family, namely
“noise”, which could also be used synonymously. However, personally I would use the word
“noise” only in the context of measurements, whereas in the context of parameter estimates
the word “uncertainty” appears to be most natural.
2.2 Errors of data

Real data is always subject to noise, i.e., there is always some uncertainty.2 The precise origin
of the noise, e.g., read-out noise, sky noise, etc., is not relevant for this manuscript. Most
methods of error estimation for parameters require knowledge about the error distribution (or
noise distribution) of the measured data. Random scatter (variance, noise) determines the
width of the error distribution, whereas the mean of the error distribution is determined by the
physical signal which may be shifted by a systematic error (bias). The easiest way to measure
the data’s error distribution is to repeat an identical measurement process and to monitor
the distribution of the results. However, this is often infeasible, e.g., because a measured
event happens rarely or because the measurement process itself is expensive (in time, money,
computational effort, etc.).
Fortunately, for many measurement processes in astronomy the mathematical type of the
error distribution is known from physical arguments, e.g., whenever spectroscopy or photometry
are carried out. In this case, the physical measurement process is counting photons in a certain
pixel. This photon counting is usually assumed to be a Poisson process, i.e., it follows the
Poisson distribution. Given a mean value of µ photons in a certain pixel, the probability of
1
Hogg et al. (2010) argue that the word “error” is not a good terminology anyway and that it should be
replaced by “uncertainty”. However, their argument assumes that systematic errors (biases) can always be
corrected.
2
One can think of real data as being one of many different “noise realisations” of the truth, but not being
the truth itself.
3
measuring n photons in this pixel is given by

µn −µ
prob(n|µ) = e . (1)
n!
Figure 1 shows examples of Poisson distributions with different mean values. In practice, the
mean µ is usually unknown and the task is to estimate the value of µ given n. We emphasise
that the mean of the Poisson distribution does not coincide with the maximum position (the
mode), since the Poisson distribution is not symmetric. The fact that the Poisson distribution
is defined for positive integers only already implies that it cannot be symmetric.
Modern instruments have high sensitivities and it is also common practice (if possible) to
choose the exposure time in the measurement such that the number of photon counts per pixel
is large. The obvious reason to do this is to improve the signal-to-noise ratio. However, there is
also another benefit: If the number of photon counts is large, the expected value of µ will also
be large. In the case of large µ, the Poisson distribution approaches the Gaussian distribution,
1 (n − µ)2

1
prob(n|µ, σ) = √ exp − , (2)
2πσ 2 2 σ2
which has much more convenient analytical properties than the Poisson distribution, as we
shall see in Sections 2.3, 3.2, and 3.3. Figure 2 shows a Poisson distribution with mean µ = 10
and a Gaussian with mean µ = 10 and variance σ 2 = µ = 10. Evidently, for a mean of only
ten photon counts per pixel, the actual Poisson distribution can be nicely approximated by a
Gaussian already. This is usually a valid approximation in the optical regime and at larger
wavelengths, whereas in the high-energy regime (UV, X-ray, gamma) it is not unheard of that
there are less than ten photon counts per pixel.
2.3 Parameter estimation

I now briefly discuss the concept of parameter estimation. Here, the focus is on model-based
estimates (e.g. estimating a mean), because they all have a common concept, whereas model-
independent parameter estimation (e.g. estimating image flux) is a very diverse field and usually
self-explanatory.
The common feature of model-based parameter estimation is that it is usually an optimi-
sation (or “fitting”) problem. The task at hand is to optimise a certain score function, e.g., to
minimise residuals or – more technically – to maximise the likelihood function 3 . For a given set
of measured data D and a given model M with parameters θ, the likelihood function is defined
via the noise model,
L(D; M, θ) = prob(D|M, θ) . (3)
In words, the likelihood is the probability of measuring the data D we have, given the model M
with parameters θ.4 Maximum-likelihood estimation then means that we choose the parameter
values such that the data we did measure were the most likely measurement outcome.5 The set
3
Bayesians prefer to maximise the posterior probability instead. I will restrict the discussion on likelihood
functions alone, since Bayesians should be happy with that assuming flat priors, whereas discussing posterior
probabilities would probably enrage some frequentists.
4
Quoting Heavens (2009): If you are confused by the conditional probabilities prob(A|B) of A given B and
prob(B|A) of B given A consider if A=pregnant and B=female. Then prob(A|B) is a few percent whereas
prob(B|A) is unity.
5
Note the subtle fact that this is actually not the question we are asking. We actually want to know what are
the most likely parameters given the data and model. Hence, we should actually maximise prob(θ|M, D) rather
than prob(D|M, θ). However, this would lead us to Bayes’ theorem and the issue of priors (e.g. see Barlow
1993) which are the cause of the long-ongoing Bayesian vs. frequentist discussion in the scientific community.
4
value xn error value xn error value xn error value xn error value xn error
7 3.18 12 3.18 12 3.08 11 2.87 8 3.32
10 3.45 9 3.14 11 3.41 7 3.07 11 2.99
11 2.92 12 3.12 13 3.32 7 2.89 9 3.08
10 3.14 13 3.03 9 3.44 12 3.06 9 3.18
8 3.43 10 3.12 10 3.31 10 2.93 9 3.40
14 2.85 11 3.07 12 3.21 6 2.90 9 3.01
Table 1: Data sample used as a standard example for all methods. All data points xn are
sampled from a Poisson distribution with mean µ = 10 (cf. Fig. 2). The columns entitled
“error” give the Gaussian standard deviations σn for each data point xn for the cases where
the error distribution is assumed to be Gaussian.
of parameters that maximise the likelihood function is called the maximum-likelihood estimate
and it is usually denoted by θ̂.
A somewhat philosopical note: Actually, the likelihood function as given by Eq. (3) is what
every parameter estimation is aiming for. This function, L(D; M, θ), contains all important
information about the data and the model, a theorem which is called likelihood principle.
However, Eq. (3) is just an abstract definition and even a more specific example (e.g. Eqs. (4)
and (5)) usually does not provide more insight. Therefore, one has to extract the information
from Eq. (3) in some way. If the model under consideration has only one or two model
parameters, it is possible to plot the likelihood function directly (e.g. Fig. 8), without involving
any optimisation procedure. Although such a plot is actually the final result of the parameter-
estimation process, people (including myself) are usually happier giving “numbers”. Moreover,
if a model has more than two parameters, the likelihood function cannot be plotted anymore.
Hence, the standard practise of encoding the information contained in the likelihood function
is by identifying the point in parameter space where the likelihood function takes its maximum
(the maximum-likelihood estimate) plus inferring the “width” of the function at its maximum
(the uncertainty). If nothing contrary is said, these two quantities usually signify the mean
value and the standard deviation of a Gaussian. Consequently, if both values are provided, one
can reconstruct the full likelihood function.
In order to “give some flesh” to the rather abstract concept of a likelihood function, two
simple examples of parameter estimation are now discussed. This allows us to see this concept
and the Poisson and Gaussian distributions “in action”. Table 1 also introduces the data sample
that will be used to demonstrate every method that is discussed using actual numbers.
2.3.1 Example 1: Estimating the mean of a Gaussian distribution

The first example is to estimate the mean µ of a sample of N measured data points D =
{x1 , x2 , . . . , xN } that are all real values. The assumption is that all data points have a Gaussian
error distribution, i.e.,
1 (xn − µ)2

1
prob(xn |µ, σn ) = p exp − , (4)
2πσn2 2 σn2
where we assume that each measurement xn has its own standard deviation σn . The probability
of all N measurements D = {x1 , x2 , . . . , xN } – the likelihood function – is (in the case of un-
correlated measurements) just the product of the probabilities of the individual measurements,
i.e.,
N
Y
L(D; µ) = prob(xn |µ, σn ) . (5)
n=1
5
There are two reasons for maximising log L instead of L. First, log L sometimes takes a much
more convenient mathematical form, enabling us to solve the maximisation problem analyti-
cally, as we shall see immediately. Second, L is a product of N potentially small numbers. If
N is large this can cause a numerical underflow in the computer. As the logarithm is a strictly
monotonic function, the maxima of L and log L will be identical. The logarithmic likelihood
function is given by
N
X
log L(D; µ) = log prob(xn |µ, σn ) . (6)
n=1
Inserting Eq. (4) yields

N
1 X (xn − µ)2
log L(D; µ) = − +C, (7)
2 n=1 σn2
where C encompasses all terms that do not depend on µ and are therefore constants during
the maximisation problem. We can now identify the sum,
N
2
X (xn − µ)2
χ = , (8)
n=1
σn2
such that log L(D; µ) = − 21 χ2 + C. In other words, maximising the likelihood function in the
case of Gaussian noise is equivalent to minimising χ2 .6 In order to estimate µ, we now take the
first derivative of Eq. (7) or Eq. (8) with respect to µ, set it equal to zero and try to solve the
resulting equation for µ. The first derivative of Eq. (7) set to zero then reads
N
d log L(D; µ) X (xn − µ)
= = 0, (9)
dµ n=1
σn2
and solving for µ yields the maximum-likelihood estimate

PN xn
n=1 σn2
µ̂ = PN 1
. (10)
n=1 σn2
This estimator is a weighted mean which underweights data points with large measurement
errors, i.e., data points that are very uncertain. For the example data set of Table 1 we get
µ̂ ≈ 10.09. This result can be simplified by assuming that all data points have identical standard
deviations, i.e., σn = σ for all xn . Our result then reads
PN
xn
µ̂ = n=1 , (11)
N
which is simply the arithmetic mean. For the example data set of Table 1 we then get µ̂ ≈ 10.07.
The derivation of the corresponding error estimation of µ̂ is postponed to Sect. 3.3.
6
Actually, we should say that minimising χ2 provides the correct estimator if and only if the error distribution
of the data is Gaussian. If the error distribution is not Gaussian, then minimising squared residuals may well
be plausible but it is not justified.
6
2.3.2 Example 2: Estimating the mean of a Poisson distribution

The second example is precisely the same task, but now the error distribution of the N mea-
surements D = {x1 , x2 , . . . , xN } (which are now all integers) should be a Poisson distribution
as given by Eq. (1). Again, we estimate the mean µ by maximising the likelihood function of
the data
N
Y
L(D; µ) = prob(xn |µ) , (12)
n=1
or rather the logarithmic likelihood function

N xn N
X µ −µ X
log L(D; µ) = log e = −N µ + log(µ) xn + C , (13)
n=1
xn ! n=1
where C again summarises all terms that do not depend on µ. Taking the first derivative with
respect to µ and setting it to zero yields
N
d log L(D; µ) 1X
= −N + xn = 0 . (14)
dµ µ n=1
Solving for µ then yields the maximum-likelihood estimate

N
1 X
µ̂ = xn , (15)
N n=1
which is the arithmetic mean, again. Do not be mistaken by the fact that the result was identical
for the Gaussian and the Poisson distribution in this example. In general, trying to estimate a
certain quantity assuming different error distributions also results in different estimators.
2.3.3 Example 3: Estimating a fraction

Finally, we also want to consider a third example which will turn out to be less well behaved
than the first two. Let us consider the situation that we are given a set of N objects, say
galaxies, of which n are in some way different from the other N − n objects, say they are active
galaxies. The task is to estimate the fraction f of special objects from these numbers. The
natural assumption here is that the likelihood function is a binomial distribution, i.e.,
n!
L(D; f ) = L(n, N ; f ) = f n (1 − f )N −n . (16)
N !(N − n)!
The desired fraction is limited to the interval f ∈ [0, 1]. What value of f maximises the
likelihood of the measured values of N and n? Let us first compute the logarithmic likelihood,
log L(D; f ) = n log f + (N − n) log(1 − f ) + C , (17)
where C contains everything that does not depend on f . The first derivative of log L w.r.t. f
reads,
∂ log L(D; f ) n N −n
= − . (18)
∂f f 1−f
Equating this to zero and solving for f yields the maximum-likelihood estimator
n
fˆ = , (19)
N
7
provided that neither n nor f nor N are zero.7 We will consider this example in two flavours:
1. N = 10 and n = 0, i.e., the given sample is very small and contains no special objects.
An error estimate enables us to assess whether this rules out the existence of these special
objects.
2. N = 30 and n = 4.
Figure 3 shows the binomial likelihood functions for both cases. An example where a binomial
distribution shows up in astronomy can, e.g., be found in Cisternas et al. (2010).
2.3.4 More general error distributions

The task of model-based parameter estimation
as outlined in the previous subsections can be
summarised as follows:
1. Identify the error distribution of the data

and write down the (logarithmic) likeli-
hood function.
2. In order to maximise the (logarithmic)

likelihood function, take the first deriva-
tive(s) with respect to the desired model
parameter(s) and set it equal to zero.
3. Try to solve the resulting (system of) equa-

tion(s) for the desired fit parameter(s).
Figure 3: Binomial likelihood functions for
The first two steps are usually not that hard. In the fraction f ∈ [0, 1] for example 3. For the
the two examples given above, step 3 was also sake of visualisation the likelihood functions
fairly simple because our model was simple (our are not normalised, which is why the y-axis
model was a constant µ) and the error distri- is unlabelled.
butions were well-behaved. However, in general
it is not possible to perform step 3 analytically
for more complex models or other error distri-
butions. In this case the optimisation has to be
done numerically. MacKay (2003) provides an overview of the most important distribution
functions that may – in principle – appear as error distributions.
Nonetheless, we should not be mistaken by the simplicity of the previous examples. Replac-
ing the error distribution measured in some experiment by an analytic distribution function,
such as a Poisson or Gaussian, is merely a makeshift. We have to keep in mind that it may
happen in practice that for some measured error distribution it may not be possible to find
such an analytic parametrisation. In such cases, things get substantially more difficult as we
can no longer write down the likelihood function. Even so, we could still work with the whole
unparametrised error distribution.
7
This result should not come as a suprise. That the fraction is estimated via n/N would have been a plausible
guess, but mind that we now have seen why it is justified.
8
2.4 Central-limit theorem

The central-limit theorem is a key result of statistics and shows up repeatedly throughout
this manuscript. In simple words, the central-limit theorem tells us that if certain regularity
conditions are met, any likelihood function is asymptotically Gaussian near its maximum. Let
us denote log L = log L(θ)~ as a function of the P parameters θ~ = {θ1 , . . . , θP } and Taylor
expand it around the maximum θ~max to second order,
2

~ ≈ log L(θ~max ) + 1 ∂ log L
log L(θ) (θ − θmax )i (θ − θmax )j . (20)
2 ∂θi ∂θj θ~max
The linear term vanishes at the maximum, be-

cause the first derivative evaluated at θ~max be-
comes zero. Consequently, log L(θ) ~ is approxi-
mately a quadratic form in θ~ near the optimum,
~
~ = elog L(θ)
i.e., L(θ) is a Gaussian. The goodness
of the Gaussian approximation depends on the
specific situation, i.e., on the measured data, its
noise, and the model used. The approximation is
good, if we are close enough. However, for error
estimation we cannot go arbitrarily close to the
maximum of the likelihood function. In fact, the
Gaussian approximation can be arbitrarily poor
in practice. Figure 4 shows an example where
this is the case. The central-limit theorem may
Figure 4: Possible failures of the central-limit be particularly problematic in case of parameters
theorem. This figure shows an example like- that are not defined on the interval (−∞, ∞) of
lihood function L as a function of some pa- a Gaussian but, e.g., are constrained on the in-
rameter θ (solid blue curves), its Gaussian terval [0, 1] such as the fraction of example 3.
approximation at the maximum (dashed red Moreover, there are non-pathologic cases where
curves), and the widths of these Gaussians the likelihood function never becomes Gaussian
(horizontal solid black lines). In panel (a) the at its maximum. For instance, consider exam-
Gaussian approximation is good and the er- ple 3.1 and Fig. 3, where the likelihood function
ror estimate reliable. In panel (b) the Gaus- even does not have a real maximum where its
sian approximation is rather poor and the first derivative would be zero. In this particular
error is substantially underestimated. case, the Taylor expansion of Eq. (20) breaks
down, because we cannot compute any deriva-
tives at the “maximum” at all. Consequently, whenever invoking the central-limit theorem,
one should carefully check the goodness of the Gaussian approximation.
2.5 Confidence intervals

In order to explain methods for error estimation, confidence intervals have to be considered.
We are all familiar with the concept of error bars, error ellipses or – in general – error contours.
They are different manifestations of confidence intervals. These confidence intervals are inferred
from the likelihood function.
2.5.1 A single parameter

Initially, confidence intervals for a single parameter are considered, in order to keep things sim-
ple. Figure 5 shows the most simple example – confidence intervals of a Gaussian distribution.
9
Figure 5: Confidence intervals for the Figure 6: Different types of 68.3% confi-
Gaussian distribution of mean hθi and dence intervals for a multimodal likelihood
standard deviation σ. If we draw N values function. The vertical dashed red line indi-
of θ from a Gaussian distribution, 68.3% cates the maximum-likelihood estimate θ̂.
of the values will be inside the interval The panels are numbered according to the
[hθi − σ, hθi + σ] as shown in panel (a), definitions in the main text.
whereas 95.5% of the values will be inside
the interval [hθi − 2σ, hθi + 2σ] as shown in
panel (b).
If we draw a sample value θ from a Gaussian with mean hθi and standard deviation σ, e.g., by
trying to estimate hθi from measured data, the deviation |θ − hθi| will be smaller than 1σ with
68.3% probability, and it will be smaller than 2σ with 95.5% probability, etc. In simple words,
if we fit some function to N data points with Gaussian errors, we have to expect that 31.7% of
all data points deviate from this fit by more than one sigma.8
The Gaussian is an almost trivial example, due to its symmetry around the mean. In
general, likelihood functions may not be symmetric and how to define confidence intervals in
these cases should be explained. For asymmetric distributions mean and maximum position do
not conincide (e.g. Poisson distribution). The actual parameter estimate θ̂ is the maximum-
likelihood estimate, i.e., it indicates the maximum of the likelihood function, not its mean. We
define the confidence interval θ− ≤ θ̂ ≤ θ+ for a given distribution function prob(θ) via (e.g.
Barlow 1993)
Z θ+
prob(θ− ≤ θ̂ ≤ θ+ ) = dθ prob(θ) = C , (21)
θ−
where usually C = 0.683 in analogy to the one-sigma-interval of the Gaussian. In practice, the
distribution function prob(θ) is usually unknown and only given as a histogram of samples of
θ. In this case, the integral in Eq. (21) reduces to the fraction of all samples θ that are between
θ− and θ+ . However, Eq. (21) does not uniquely define the confidence interval, an additional
criterion is required. Possible criteria are (e.g. Barlow 1993):
1. Symmetric interval: θ− and θ+ are symmetric around the parameter estimate, i.e., θ̂−θ− =
θ+ − θ̂.
2. Shortest interval: θ+ − θ− is smallest for all intervals that satisfy Eq. (21).
8
If you are presented fitted data where the fit goes through all 1σ-errorbars, you should definitely be sceptical.
10
P 1 2 3 4 5 6 7 8 9 10
cP 0.6827 0.3935 0.1988 0.1149 0.0715 0.0466 0.0314 0.0217 0.0152 0.0109
Table 2: Confidence cP contained within a 1σ contour of a P -dimensional Gaussian as given by

Eq. (23). In ten dimensions, the 1σ contour contains a ridiculously small confidence of ≈ 1.1%.
R θ−
3. Central interval: The probabilities above and below the interval are equal, i.e., −∞
dθ prob(θ) =
R∞
θ+
dθ prob(θ) = (1 − C)/2.
In case of a symmetric distribution, e.g., a Gaussian, all three definitions are indeed equivalent.
However, in general they lead to different confidence intervals. Figure 6 shows the 68.3%
confidence intervals9 resulting from the three definitions for an example distribution that could
be a likelihood function resulting from a parameter estimate. In practice, there is usually no
preference for any of these definitions10 , it should only be made explicitly clear which one is
used.
2.5.2 Two or more parameters

If we are estimating two or more parameters and are interested in estimating the joint confidence
region, things become considerably more difficult. This difficulty largely stems from the fact
that multiple parameters will usually exhibit mutual correlations. The following discussion
largely follows Barlow (1993).
First, consider the case where the central-limit theorem indeed ensures that some (mul-
tidimensional) likelihood function is approximately Gaussian at its maximum position. Such
(multivariate) Gaussians, with P -dimensional mean vector µ ~ and P × P covariance matrix Σ,

1 1 T −1
prob(~x|~µ, Σ) = p exp − (~x − µ~ ) · Σ · (~x − µ ~) , (22)
(2π)P det Σ 2
provide ellipsoidal error contours, i.e., they are capable of describing linear correlations in the
parameters, but not nonlinear correlations such as “banana-shaped” error contours. However,
even in this simple case, things are complicated. The reason for this is that the one-sigma
contour no longer marks a 68.3% confidence region as in Fig. 5. It is straight-forward to
compute that the one-sigma contour of a two-dimensional Gaussian marks a 39.4% confidence
region, whereas in three dimensions it is just a 19.9% confidence region.11 In general, the
confidence cP contained inside a one-sigma contour of a P -dimensional Gaussian with P > 1 is
given by,
Z 1
1 P −2 2
cP = P/2
2π 2 dr rP −1 er /2 . (23)
(2π) 0
Table 2 gives cP for P -dimensional Gaussians with P ≤ 10, in order to give an impression of
how quickly the confidence declines. Evidently, one needs to be very careful when interpreting
one-sigma contours in more than one dimension.
9
Note our terminology: We are talking of a “68.3% confidence interval”, not of a “one-sigma interval”.
10
A symmetric confidence interval may not be sensible in case of a highly asymmetric likelihood function.
As a nice example, consider the likelihood function of example 3.1 shown in Fig. 3. Furthermore, the central
interval would cause the “maximum” at fˆ = 0 to lie outside this confidence interval. h i
11
R 2π Rσ 1 r2
In order to obtain the two-dimensional result, solve the integral 0 dϕ 0 dr r 2πσ 2 exp − 2σ 2 that assumes
a spherically symmetric Gaussian given in polar coordinates.
11
If the central-limit theorem does not apply – e.g., because the number N of measured data
is small or the likelihood function itself is not well-behaved – things get even more involved.
Nonlinear correlations in the parameters, i.e., “banana-shaped” error contours, are an obvious
indicator for this case. The symmetric confidence region can still be defined easily, but it
obviously lacks the ability to describe parameter correlations. Identifying the “shortest region”
or “central region” may be computationally very expensive. Barlow (1993) recommends a
definition of the confidence region via the contour at which the logarithmic likelihood is 0.5
lower than that at its maximum, i.e., where the likelihood function takes 1/e of its maximum
value. However, the degree of confidence of the resulting region strongly depends on the number
of dimensions, similarly to the Gaussian case discussed above.
As an alternative approach I now discuss a
method that is designed to provide a 68.3% con-
fidence region. Similarly to the method recom-
mended by Barlow (1993), it employs contours
on which the likelihood function is constant. The
recipe is as follows:
1. Locate the maximum of the likeli-

hood function, Lmax , i.e., perform the
maximum-likelihood estimation of the pa-
rameters.
2. Identify the contours where the likelihood

function takes some constant value L0 <
Lmax and integrate L over these regions in
Figure 7: Confidence regions for an example
order to get their confidence level.
likelihood function resulting from Barlow’s
3. Adjust the contour level L0 such that the method (a) and the alternative method pre-
resulting region in parameter space has sented in this manuscript (b). The horizon-
68.3% confidence. tal dashed red lines indicate the slices at L0 .
In panel (a) Barlow’s region has 80.4% con-
We visualise this method in Fig. 7 where, for fidence. In panel (b) the 68.3% confidence
the sake of visualisation, there is only a sin- region consists of two regions that are not
gle parameter. Moreover, Fig. 7 also demon- connected.
strates that this method may provide a 68.3%
confidence region that consists of various dis-
connected regions, if the likelihood function is
multimodal. Actually, this is not a real problem
because the disconnected regions then also indicate the other local maxima. If the likelihood
function is highly multimodal as in Fig. 7 and also in Fig. 6, this clearly makes more sense
than using confidence intervals that are symmetric, central or shortest.
2.5.3 The worst-case scenario

A final word of caution, before specific methods are discussed: As discussed earlier, parameter
estimation via maximising a likelihood function and quantifying its uncertainties via certain
intervals is just a makeshift. If it is possible, it greatly simplifies the interpretation because
it describes the likelihood function by a simple set of numbers. However, in practice, this
procedure may turn out to be infeasible. In such cases, we have to resort to working with the
full likelihood function itself, without further simplifications.
12
3 Error estimation for model-based parameters

I now discuss how to perform an error estimation in the case of model-based parameter esti-
mates. In particular, I discuss four different methods – brute-force grids, varying χ2 , the Fisher
matrix, and Monte-Carlo methods – which are very common in astronomy.
3.1 Brute-force grids

A very straight-forward approach to both pa-
rameter estimation and subsequent error esti-
mation are brute-force grids. In this case, we
span a (rectangular) grid in the parameter space
and evaluate the likelihood function at every grid
point. If we have only one or two model param-
eters, we can then plot the likelihood function
directly and we can also directly infer the error
estimates from the contour lines.
This method is very robust, because it only
relies on the assumption that the error distribu-
tion of the measured data is correct, i.e., that we
are certain about the likelihood function. How-
ever, if the number of model parameters is large,
this method becomes computationally infeasible. Figure 8: Error estimation via brute-force
Computational infeasibility may even be an issue grids for the standard data. Panel (a) as-
for only two or even a single model parameter, sumes the error distribution to be Poisso-
if it is very expensive to evaluate the likelihood nian, whereas panel (b) assumes it to be
function. Gaussian. Both likelihood functions L(µ)
(solid blue curves) peak at µ ≈ 10.07. The
3.1.1 Application to example data dashed red curves are matched Gaussians
that are slightly shifted upwards for the sake
This method is applied to the example data of of visibility. The Gaussian’s standard devia-
Table 1 and used to try to estimate the mean and tions (which provide the error estimates) are
its error. For candidate values of µ ∈ [1, 19] the σ̂µ ≈ 0.59 for panel (a) and σ̂µ ≈ 0.59 for
likelihood function L(µ) is computed assuming panel (b).
the data’s error distribution to be Poisson and
Gaussian. Figure 8 shows L(µ) for both cases.
The function L(µ) is then matched by a Gaus-
sian12 and the standard deviation of this Gaussian is the error estimate. This provides the
estimate σ̂µ ≈ 0.59 for both Poisson and Gaussian error distributions. Concerning example 3,
Figure 3 is essentially a brute-force grid parameter estimation.
3.2 Varying χ2
χ2 has already been introduced in Eq. (8). Let us assume that the model parameters were
estimated by minimising χ2 . We then vary the model parameters slightly around the optimal
values such that χ2 changes by less than 1. In other words, we look for the contour where
χ2 = χ2min + 1, thereby defining the error estimate of the model parameters. The basic idea
here is that if the likelihood function were Gaussian, this would yield the 1σ contour.
12
This is motivated by the central-limit theorem, but the Gaussian approximation needs to be checked.
13
The crucial assumption of this method is that the error distribution of the measured data
is indeed Gaussian, because otherwise using a χ2 does not make sense (Sect. 2.3.1). Moreover,
this method relies on an accurate measurement of the data errors σn .
3.2.1 An example of bad practice

In this section, I want to discuss an example of bad practise that I discovered as a refereed
publication, which provides a couple of lessons to learn. For obvious reasons, I will not give a
reference here. The publication pushes the method of varying χ2 even further in an attempt to
ensure that the measurement errors σn are correct. Fitting a model with P parameters to N
data points, the authors demand that
χ2
χ2 = N − P ⇔ χ2red = = 1, (24)
N −P
where χ2red is called “reduced χ2 ”. Loosely speaking, this means that each degree of freedom
contributes one standard deviation. The authors then try to correct for potentially wrong
values of the σn by rescaling them such that χ2red = 1.
There are several objections to this method.
In order to see this, we need to consider the as-
sumptions involved:
1. The error distribution has to be Gaussian,
as in varying χ2 . The authors do not jus-
tify this assumption.
2. The model has to be linear in all P fit

parameters. If the model is nonlinear, we
cannot demand that χ2red = 1, because the
derivation of χ2 = N − P implicitly as-
sumes linearity in all parameters (e.g. cf.
Barlow 1993).13 Again, the authors do not
comment on this, in fact, they even never
explicitly say what their model is. Figure 9: Error estimation via varying χ2 for
the standard data. The minimum of χ2 oc-
2
3. By demanding χred = 1, we explicitly claim curs at µ̂ ≈ 10.09 as suggested by Eq. (10).
that the model we are using is the correct There χ2 ≈ 12.4, as indicated by the lower
model that was underlying the data. This horizontal, dashed, red line. The upper hor-
is a rather optimistic claim. Of course, we izontal, dashed, red line indicates the value
would like our model to be the truth, but 12.4 + 1.0 = 13.4. The error estimates are
this claim requires justification. Moreover, defined by those points where χ2 = 13.4, as
in many practical situations we are actu- indicated by the two vertical solid black lines.
ally not interested in the truth, but rather The resulting error estimate is σ̂µ ≈ 0.57.
make use of a model which expresses a rea-
sonable simplification.
Even if all these assumptions are met, the
method is in fact only applicable if the degrees of freedom N − P are large. The reason is
that the uncertainty in the measured data does not only cause an uncertainty in the model
13
More precisely: We need to know the number of degrees of freedom in order to evaluate χ2red . As we
are going to show in an upcoming publication, estimating the number of degrees of freedom is possible but
nontrivial for linear models. For nonlinear models, however, estimating the number of degrees of freedom is
virtually impossible, unless the central-limit theorem provides a good approximation.
14
parameters, but also an uncertainty in the value of χ2 itself. The value of χ2 is subject to
the so-called χ2 -distribution (e.g. see Barlow 1993) whose expectation value is indeed N − P .
However, this distribution is not sharp but has a nonzero variance of 2(N − P ). Consequently,
if N − P is small, there is a large relative uncertainty on the value of χ2 . This means χ2 may
deviate substantially from N − P even though the model is linear and correct.
3.2.2 Application to example data

In Figure 9, this method is used to estimate the error on the mean estimator given by Eq. (10).
Given the data of Table 1 and µ̂ ≈ 10.09, the minimal value of χ2 is ≈ 12.4.14 The next step is
to look for those values of µ where χ2 takes the values 12.4 + 1.0 = 13.4. In this simple case,
these two points are symmetric around the minimum because the model is linear. The resulting
error estimate is σ̂µ ≈ 0.57 as in Sect. 3.1.1. For more general models, the error estimates can
be asymmetric. This method is not applicable to example 3, because the underlying error
distribution is not Gaussian.
3.3 Fisher matrix

Error estimation via the Fisher matrix is a very popular approach (e.g. see Heavens 2009).
Let us assume we have fitted our model parameters to our data via maximising the logarith-
mic likelihood (e.g. minimising χ2 in case of Gaussian noise). This method is based on the
central-limit theorem which tells us that any well-behaved likelihood function is asymptotically
Gaussian near its maximum. We may therefore write

1 ~ ~ T
L(θ) ∝ exp − (θ − θ0 ) · Σ · (θ~ − θ~0 ) ,
−1
(25)
2
which is a P -dimensional (P -variate) Gaussian with mean θ~0 and covariance matrix Σ. This
covariance matrix is the desired error estimate. On the diagonals it contains the variance
estimates of each individual θp and the off-diagonals are the estimates of the covariances (see
e.g. Barlow 1993, for more information about covariances). Comparing Eqs. (20) and (25), we
identify
2 −1
∂ log L
Σ̂ = − . (26)
∂θi ∂θj
Care should be taken with the order of indices and matrix inversion. The matrix of second
derivatives of log L is called “Fisher matrix”. If the second derivatives of log L can be evaluated
analytically, this method may be extremely fast from a computational point of view. However,
if this is impossible, they can usually also be evaluated numerically. By construction, this
method can only describe elliptical error contours. It is impossible to obtain “banana-shaped”
error contours from this method.
Of course, this method also invokes assumptions that have to be checked. Those assumptions
are:
1. The error distribution of the measurements is known, i.e., the likelihood function is defined
correctly.
2. The second-order Taylor expansion of Eq. (20) is a good approximation.

14
Note that this deviates from N − 1 = 29 due to uncertainty in χ2 induced by the small number of data
points.
15
This second assumption is the actual problem. Although the central-limit theorem ensures this
asymptotically, Fig. 4 shows an example where this assumption of the Fisher matrix breaks
down. There are two simple tests to check the validity of the resulting covariance-matrix
candidate Σ̂. A valid covariance matrix has to be positive definite, i.e., ~xT · Σ̂ · ~x > 0 for any
nonzero vector ~x, and both tests try to check this:
1. Compute the determinant det Σ̂. If det Σ̂ ≤ 0, Σ̂ is not valid.
2. Diagonalise the matrix Σ̂ in order to determine its eigenvalues. If any eigenvalue is

negative or zero, Σ̂ is not valid.
The first test is usually easier to perform, whereas the second test is more restrictive. It is
strongly recommended that these tests are applied whenever this method is used. Unfortu-
nately, these tests are only rule-out criteria. If Σ̂ fails any of these tests, it is clearly ruled
out. However, if it passes both tests, we still cannot be sure that Σ̂ is a good approximation,
i.e., that the Gaussian is indeed a decent approximation to the likelihood function at its maxi-
mum. Nevertheless, the major advantage of this method is that it is very fast and efficient, in
particular if we can evaluate the second derivatives of log L analytically.
There are also situations where the Fisher matrix is definitely correct. This is the case for
Gaussian measurement errors and linear models. In this case L(θ) ~ is truly a Gaussian even
2
without any approximation. For instances, inspect Eq. (8): χ is a quadratic function of µ,
2
and since L(µ) ∝ e−χ /2 the likelihood function is a Gaussian.
3.3.1 Example 1 revisited

In order to see the Fisher matrix “in action”, we now return to example 1 from Sect. 2.3.1.
The task was to estimate the mean µ of N data points {x1 , x2 , . . . , xN } that are drawn from a
Gaussian error distribution. The maximum-likelihood estimator for µ is given by Eq. (7). The
task now is to use the Fisher matrix in order to derive the error estimate for Eq. (7). Equation
(9) is derived once more with respect to µ,
N
d2 log L(D; µ) X 1
= . (27)
dµ2 σ2
n=1 n
For the error estimate of µ̂ this then yields

N
!−1
X 1
σ̂µ2 = . (28)
σ2
n=1 n
If all N data points are again assumed to have identical errors, σn = σ, this simplifies to
σ̂µ2 = σ 2 /N which is the error estimate for the arithmetic mean in case of Gaussian measurement
errors.

The same exercise is repeated for example 2, where the error distribution is assumed to be
Poissonian. The second derivative of Eq. (14) with respect to µ reads
N
d2 log L(D; µ) 1 X
= − xn . (29)
dµ2 µ2 n=1
16
For the error estimate of µ̂ this yields
µ̂2 µ̂
σ̂µ2 = PN = , (30)
n=1 xn
N
PN
where n=1 xn = N µ̂ has been identified according to the estimator of Eq. (15).

Inserting the example data of Table 1 into Eq. (28), an error estimate of σ̂µ ≈ 0.57 is obtained,
which is in agreement with the error estimates from Sections 3.1.1 and 3.2.2. In the case of
a Poisson error distribution, insertion of the example data of Table 1 into Eq. (30) yields
σ̂µ ≈ 0.58, which is also in agreement with the result of Sect. 3.1.1. Both results are reliable,
since in both cases the likelihood function is well approximated by a Gaussian, as we have seen
in Fig. 8.

Let us compute the second derivative of the logarithmic likelihood given by Eq. (17) w.r.t. f .
We obtain,
d2 log L n N −n N3
= − − = − , (31)
df 2 f 2 (1 − f )2 n(N − n)
where we have inserted the maximum-likelihood estimator fˆ = n/N . Hence, the error estimate
pretends to be
n(N − n)
σf2 = ,
N3
which yields σf2 = 0 in example 3.1 and σf2 ≈ 0.004 in example 3.2. Now we are lucky, because
σf2 = 0 tells us that we forgot that in example 3 the likelihood function is binomial and not
Gaussian, i.e., this whole calculation was nonsense. Unfortunately, this is not necessarily that
obvious, as the result for example 3.2 shows.
3.4 Monte-Carlo methods

Monte-Carlo methods directly draw samples from the likelihood function, i.e., they guess values
of the fit parameters and accept them with the probability defined by the corresponding value
of the likelihood function. Although these methods may appear difficult at first glance, Monte-
Carlo sampling is actually the most intuitive approach to error estimation. The reason is its
similarity to measuring errors of data by repeating the measurement process and monitoring
the results. The strength of all Monte-Carlo methods is that they use a minimum amount of
assumptions. Their sole and only assumption is that the error distribution of the measured
data is known correctly, i.e., that we are certain about the likelihood function. There are no
further requirements such as Gaussianity, which renders this approach very general.
There are many different types of Monte-Carlo methods for many different situations. It
would be beyond the scope of this manuscript to explain these methods. Instead, I explain
which methods are useful in which kinds of situations and refer the interested reader to the
literature. MacKay (2003), for example, provides an excellent introduction to all the methods
named here.
17
The first criterion in choosing a particular

Monte-Carlo algorithm is the number of model
parameters. If the number of parameters is small
– say 1 or 2 – we can use one of the following
Monte-Carlo algorithms: uniform sampling, im-
portance sampling, or rejection sampling. How-
ever, if the number of parameters is large, these
algorithms quickly become inefficient, i.e., com-
putationally expensive. In the case of many
model parameters, there is the family of Markov-
chain Monte-Carlo (MCMC) algorithms, whose
computation times scale linearly with the num-
ber of model parameters. The most famous (and
most simple) type of MCMC algorithm is the Figure 10: Error estimation via Monte-Carlo
Metropolis-Hastings algorithm. However, this sampling for the standard data. The distri-
algorithm involves a stepsize for each parame- butions of µ resulting from the MCMC algo-
ter that needs to be fine tuned by hand. This is rithm assuming Poisson errors in panel (a)
a real problem in the case of many parameters. and Gaussian errors in panel (b) are both
We therefore recommend the slice-sampling al- well approximated by a Gaussian. In panel
gorithm, which also involves a stepsize but it is (a) the Gaussian is given by µ̂ = 10.10 and
tuned automatically without human interaction. σ̂µ = 0.58. In panel (b) the Gaussian is given
In fact, there is also Gibbs sampling which does by µ̂ = 10.10 and σ̂µ = 0.57.
not involve any stepsizes at all. However, Gibbs
sampling can only be used directly for trivial models. In order to employ Gibbs sampling in a
wider context, one has to combine it with other one-dimensional sampling algorithms.
In practice, we start an MCMC algorithm at the estimated maximum θ~0 of the likelihood
function and let it draw M samples of model parameters {θ~1 , . . . , θ~M } from the likelihood
function.15 All we then need to do is to analyse the distribution of this set of parameters.16
For instance, assuming Gaussianity, we can do this via estimating the mean and the covariance
of this sample. However, the distribution can also be analysed more generally such that it is
possible to describe nonlinear dependencies between different model parameters, e.g., “banana
shaped” error contours.

Once again, this method is applied to the example data given in Table 1. A Metropolis-
Hastings algorithm is used with stepsizes inspired by the error estimates obtained previously.
For likelihood functions assuming Poissonian and Gaussian error distributions, respectively, the
MCMC algorithm is iterated 100,000 times. Afterwards, the resulting sequences are thinned
by picking out every tenth data point and discarding everything else. From the remaining
10,000 data points the arithmetic mean and its standard deviation are estimated. In Fig. 10,
the resulting data distribution is overplotted by a corresponding Gaussian, which provides the
error estimate. For the Poisson error distribution σ̂µ = 0.58 is obtained, and σ̂µ = 0.57 for
Gaussian error distribution. This is again in agreement with previous results.
15
Initialising the MCMC at the optimum θ~0 also has the advantage that we do not need to use any convergence
diagnostics, since we are already at the optimum. Convergence diagnostics for MCMC algorithms are still an
open issue (Cowles & Carlin 1996).
16
In practice we often need to “thin” this sequence due to autocorrelations in the Markov chain, e.g., by
selecting every tenth point and discarding all others.
18
4 Error estimation for model-independent parameters

So far we have focused our attention on error estimation for model-based parameters. Model-
based inference usually involves an optimisation problem which can complicate matters. There-
fore, model-independent parameters that usually do not involve any optimisation are very
popular in astronomy. I now explain how to estimate errors for parameters of this type.
4.1 Error propagation

One simple method is error propagation. If we can express the parameter as a function of the
measured data and if the error distribution of the data is Gaussian, we can employ Gaussian
error propagation.
To give an example, let us consider the flux F of a galaxy. If the image background has
been subtracted correctly, the flux is just given by the sum of all N pixel values fi in the image,
N
X
F̂ = fi . (32)
i=1
As argued in Sect. 2.2, the error distribution in photometric images is in excellent approxima-
tion to Gaussian, if the exposure time was long enough. If we denote the measurement error
of pixel i by σi , we can then estimate the error of F̂ via
N
!2 N
2
X ∂ F̂ X
σ̂F = σi = σi2 , (33)
i=1
∂f i i=1
which is fairly simple in this case. However, this can become very cumbersome for more general
model-independent parameters. In particular, it is impossible if a model-independent parameter
involves an operation on the measured data that is not differentiable, e.g., selecting certain data
points. In fact, error propagation can also be applied to model-based parameter estimates, if
these estimators can be expressed as differentiable functions of the data, e.g., as is the case
for linear models with Gaussian measurement errors. For instance, Equation (28) can also be
derived from Eq. (10) using this method, giving the same result for the example data of Table
1. However, in general, this is not the case.
4.2 Resampling the data

A very elegant method for error estimation is to resample the measured data (for an example
see e.g. Burtscher et al. 2009). Again, the assumption is that we know the correct error distri-
bution of the measured data. For the sake of simplicity, let us assume this error distribution
is Gaussian. For each data point xn we invoke a Gaussian with mean xn and standard devi-
ation σn as measured during the experiment. We can now sample a new data point x0n from
this distribution.17 Doing this for all pixels, we get an “alternative” noise realisation of the
measurement, e.g., a new image that can be interpreted as an alternative measurement result.
We then estimate our model-independent parameter for this resampled data and the result will
differ slightly from that obtained from the actual data. Repeating this resampling process,
e.g., 100 times, and monitoring the resulting parameter estimates, we get a distribution of this
model-independent parameter from which we can then infer an upper limit for the uncertainty.
For instance, if the resulting parameter distribution is approximately Gaussian, the upper limit
will be given by the Gaussian’s standard deviation.
17
In fact, this method is a Monte-Carlo method, too.
19
Figure 11: Error estimation via resampling Figure 12: Error estimation via bootstrap-
the data, using the standard data. The ping the data, using the standard data.
distributions of µ resulting from the re- The distributions of µ result from the boot-
sampling procedure assuming Poisson er- strapping procedure and estimating the
rors in panel (a) and Gaussian errors in Poisson mean via Eq. (15) in panel (a)
panel (b) are both well approximated by and estimating the Gaussian mean via Eq.
a Gaussian. In panel (a) the Gaussian is (10) in panel (b). Both distributions are
given by µ̂ = 10.05 and σ̂µ = 0.57. In panel well approximated by a Gaussian. In panel
(b) the Gaussian is given by µ̂ = 10.11 and (a) the Gaussian is given by µ̂ = 10.06 and
σ̂µ = 0.55. σ̂µ = 0.36. In panel (b) the Gaussian is
given by µ̂ = 10.08 and σ̂µ = 0.38.
Why does this method only provide an upper limit for the uncertainty? The reason is
that even though we are using the correct error distribution of the measured data, we are
centering this error distribution at the measured value instead of the (unknown) true value.
This introduces additional scatter and leads us to overestimate the uncertainty. Nevertheless,
it may still be acceptable to use the overestimated uncertainty as a conservative estimate,
depending on the precise scientific question.
This method is a very intuitive approach to error estimation, because it simulates repeated
measurements of the data. It can also be applied to error estimation for model-based parameter
estimates.

Again, this method is applied to the example data of Table 1. The measured data points
are assumed to be the means of the error distribution (Poisson and Gaussian) and each data
point is resampled from its error distribution. For this resampled data, the Poisson mean is
estimated via Eq. (15) and the Gaussian mean via Eq. (10). This is repeated 1,000 times and
the estimated means are monitored. Figure 11 shows the resulting distribution of mean values.
The error estimates agree well with those obtained from the other methods.
4.3 Bootstrapping
Bootstrapping (Efron 1979; Hastie et al. 2009) is another resampling method for error estimation
that can be applied to model-based as well as model-independent parameter estimation. Let
us assume we have N measurements {x1 , x2 , . . . , xN } from which we estimate some parameter.
Again, we resample our data in order to create “alternative” data sets from which we then
repeatedly estimate the parameter of interest, monitoring its distribution as before. However,
20
the details of the resampling process are different from those in Sect. 4.2. Instead of resampling
each data point from its individual error distribution, we draw new samples from the measured
data set itself. Drawing these “bootstrap” samples is done with replacement, i.e., the same
data point can occur multiple times in our bootstrap sample. To give an example, we consider
a data set {x1 , x2 , x3 , x4 }. Some examples for its bootstrap samples are:
• {x1 , x2 , x3 , x4 } itself, of course.
• {x1 , x2 , x1 , x4 },
• {x1 , x2 , x2 , x4 },
• {x1 , x3 , x3 , x3 },
• {x2 , x2 , x2 , x2 }, which is bad but possible.
As the same data point can occur multiple times but ordering is not important, for N data
points the total number of possible different bootstrap samples is

2N − 1 (2N − 1)!
= . (34)
N N !(N − 1)!
In practice, the number of bootstrap samples chosen is set to some useful number, where
“useful” is determined by the trade-off between computational effort and the desire to have as
many samples as possible in order to get a good estimate.
The major advantage of bootstrapping is that the error distribution of the measured data
does not need to be known (unless the parameter estimation itself requires it). The crucial
assumption here is that the measured data sample itself encodes the information about its
error distribution. However, the parameter estimation must be capable of dealing with the
bootstrapped samples which, in general, include certain data points multiple times while com-
pletely lacking other data points. For instance, the flux estimator of Eq. (32) would not be
capable of handling bootstrap samples, since all pixels have to contribute precisely once. Nev-
ertheless, if we know the data’s error distribution, we should really exploit this knowledge by
using, e.g., resampling instead of bootstrapping.

Finally, also bootstrapping is applied to the data of Table 1. This data sample contains N = 30
values, i.e., the total number of possible bootstrap samples is ≈ 5.9·1016 . Then, 10,000 bootstrap
samples are drawn and for every sample the Poisson mean is estimated via Eq. (15) and the
Gaussian mean via Eq. (10). Figure 12 shows the resulting distributions. Obviously, the mean
values are estimated correctly. However, the errors are underestimated, especially in the case
of the Gaussian. The likely explanation is that the data sample is not large enough in order to
contain sufficient information about its underlying error distribution.
5 Propagating measurement errors through data-reduction

pipelines
Usually, directly measured data has to be preprocessed before it can be analysed further. The
preprocessing is typically done by some data-reduction pipeline. For instances, the data coming
out of a spectrograph is in its raw state rather useless. Before one can analyse it, one has to
subtract the bias, correct for the flat field response of the CCD, and calibrate wavelength and
21
flux. This preprocessing is usually done either using complete instrument-specific pipelines,
more general software routines such as those found in IRAF or MIDAS, or a combination of
both. Now the question arises, how to propagate the errors on the initial measurements through
such complex preprocessing?
In general, an analytic error propagation such as that discussed in Sect. 4.1 is impossible,
either because the data-reduction pipeline is too complex, or because the pipeline is used as
a “black box”. Nevertheless, it is possible to propagate the errors through the pipeline via
resampling (Sect. 4.2), though it may be computationally expensive. Let us outline this
method for our previous example of spectral measurements. We do have the raw spectral data,
the measured bias fields and flat fields. We resample each of these fields as described in Sect.
4.2, say N resamplings, assuming the measurement errors are Gaussian (or Poisson, if we count
only few photons). Then, we feed each resampling instance through the pipeline and monitor
the outcome. The result will be a set of reduced spectra, which provide an estimate of the
reduced spectrum’s error distribution.
Of course, this method may be computationally expensive in practice.18 However, as we
have argued earlier, an error estimate is inevitably necessary. Therefore, if this method is the
only possibility to get such an error estimate, computational cost is not an argument.19
6 Summary
I have discussed different methods for error estimation that apply to model-based as well as
model-independent parameter estimates. The methods have been briefly outlined and their
assumptions have been made explicit. Whenever employing one of these methods, all assump-
tions should be checked. Table 3 summarises all the methods discussed here and provides a
brief overview of their applicability. It was beyond the scope of this manuscript to describe
all methods in detail. Where possible I pointed to the literature for examples. Furthermore, I
have also outlined how one can propagate errors through data-reduction pipelines.
My recommendations for error estimation are to use Monte-Carlo methods in case of model-
based parameter estimates and Monte-Carlo-like resampling of the measured data in case of
model-independent parameter estimates. These methods only require knowledge of the mea-
surement errors but do not invoke further assumptions such as Gaussianity of the likelihood
function near its maximum. Bootstrapping may also be an option if sufficient data are available.
I conclude with some recommendations for further reading:
• Barlow (1993): An easy-to-read introduction into the basics of statistics without going
too much into depth. Recommendable to get a first idea about parameter estimation,
error estimation, and the interpretation of uncertainties.
• Press et al. (2002): This book contains some very useful chapters about statistical theory,
e.g., linear least-squares fitting. It is excellent for looking up how a certain method works,
but it is not meant as an introduction to statistics.
• Hastie et al. (2009): A good textbook giving a profound introduction into analysing data.
However, the main focus of this book is on classification problems, rather than regression
problems. Given the rather mathematical notation and compactness, this book requires
a certain level of prior knowledge.
18
Although after the necessary groundwork has been covered, the batch processing of simple spectra usually
takes no more than a few seconds a piece.
19
In fact, one may argue that computational cost is not an argument anyway. If something is computationally
too expensive, then one should buy more computers! Unfortunately, this approach is usually hampered by the
fact that licensed software is very popular in astronomy (e.g. IDL).
22
method section model-based model-independent data error contours

brute-force grids 3.1 yes no known arbitrary
varying χ2 3.2 yes no Gaussian arbitrary
Fisher matrix 3.3 yes no known elliptical
Monte-Carlo methods 3.4 yes no known arbitrary
error propagation 4.1 depends yes Gaussian elliptical
resampling the data 4.2 yes yes known arbitrary
bootstrapping 4.3 yes yes unknown arbitrary
Table 3: Methods for error estimation discussed in this manuscript. This table gives a brief
overview of each method: Specifically, whether a certain method is applicable to model-based
and/or model-independent parameter estimates, whether knowledge about the data’s error
distribution is necessary, and what kind of error contours can be estimated.
• MacKay (2003): The focus of this textbook is again mainly on classification, but it
also provides a broader overview of concepts of data analysis and contains an excellent
introduction to Monte-Carlo methods. This textbook also requires a certain level of prior
knowledge.
Acknowledgements I would like to thank David W. Hogg for reading this manuscript and
providing valuable comments on the technical issues. Furthermore, I would like to thank
Ellen Andrae and Katherine Inskip for helping me to get this rather technical subject down to
something that is hopefully readable. Katherine also helped me to get the examples with spec-
troscopy right. Last but not least, I want to thank Ada Nebot Gomez-Morán, who complained
that she was dearly missing a guide to error estimation during her PhD. I hope she approves
of this manuscript. This is the first revised version of the original manuscript.
References
Barlow, R. 1993, Statistics: A Guide to the Use of Statistical Methods in the Physical Sciences
(Wiley VCH)
Burtscher, L., Jaffe, W., Raban, D., et al. 2009, ApJL, 705, L53
Cisternas, M., Jahnke, K., Inskip, K. J., et al. 2010, ArXiv e-prints
Cowles, M. K. & Carlin, B. P. 1996, Journal of the American Statistical Association, 91, 883
Efron, B. 1979, Ann. Statist., 7, 1
Hastie, T., Tibshirani, R., & Friedman, J. 2009, The Elements of Statistical Learning: Data
Mining, Inference, and Prediction. (Springer-Verlag)
Heavens, A. 2009, ArXiv e-prints
Hogg, D. W., Bovy, J., & Lang, D. 2010, ArXiv e-prints
MacKay, D. 2003, Information Theory, Inference, and Learning Algorithms (Cambridge Uni-
versity Press)
Press, W., Teukolsky, S., Vetterling, W., & Flanner, B. 2002, Numerical Recipes in C++
23

Error Estimation in Astronomy: A Guide

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Error Estimation in Astronomy: A Guide

Uploaded by

Copyright:

Available Formats

Error estimation in astronomy: A guide

Estimating errors is a crucial part of any scientific analysis. Whenever a parameter is

• Random errors during the measurement process.

• Systematic errors during the measurement process.

• Systematic errors introduced by a model.

Figure 1: Examples of Poisson distribu- Figure 2: Poisson distribution with µ = 10

2.2 Errors of data

measuring n photons in this pixel is given by

2.3 Parameter estimation

2.3.1 Example 1: Estimating the mean of a Gaussian distribution

Inserting Eq. (4) yields

and solving for µ yields the maximum-likelihood estimate

2.3.2 Example 2: Estimating the mean of a Poisson distribution

or rather the logarithmic likelihood function

Solving for µ then yields the maximum-likelihood estimate

2.3.3 Example 3: Estimating a fraction

log L(D; f ) = n log f + (N − n) log(1 − f ) + C , (17)

2.3.4 More general error distributions

1. Identify the error distribution of the data

2. In order to maximise the (logarithmic)

3. Try to solve the resulting (system of) equa-

2.4 Central-limit theorem

The linear term vanishes at the maximum, be-

2.5 Confidence intervals

2.5.1 A single parameter

Table 2: Confidence cP contained within a 1σ contour of a P -dimensional Gaussian as given by

2.5.2 Two or more parameters

1. Locate the maximum of the likeli-

2. Identify the contours where the likelihood

2.5.3 The worst-case scenario

3 Error estimation for model-based parameters

3.1 Brute-force grids

3.2.1 An example of bad practice

2. The model has to be linear in all P fit

3.2.2 Application to example data

3.3 Fisher matrix

2. The second-order Taylor expansion of Eq. (20) is a good approximation.

1. Compute the determinant det Σ̂. If det Σ̂ ≤ 0, Σ̂ is not valid.

2. Diagonalise the matrix Σ̂ in order to determine its eigenvalues. If any eigenvalue is

3.3.1 Example 1 revisited

For the error estimate of µ̂ this then yields

3.3.2 Example 2 revisited

For the error estimate of µ̂ this yields

3.3.3 Application to example data

3.3.4 Example 3 revisited

3.4 Monte-Carlo methods

The first criterion in choosing a particular

3.4.1 Application to example data

4 Error estimation for model-independent parameters

4.1 Error propagation

4.2 Resampling the data

4.2.1 Application to example data

• {x1 , x2 , x3 , x4 } itself, of course.

• {x2 , x2 , x2 , x2 }, which is bad but possible.

4.3.1 Application to example data

5 Propagating measurement errors through data-reduction

method section model-based model-independent data error contours

Efron, B. 1979, Ann. Statist., 7, 1

Heavens, A. 2009, ArXiv e-prints

Hogg, D. W., Bovy, J., & Lang, D. 2010, ArXiv e-prints

You might also like